Correct the Principle-1 over-read in the em-graph header comment (fixes #587)

[lantisprime]

Fixes #587.

scripts/em-graph.mjs:29-31 cited Principle 1 to justify not persisting a derived graph index. That over-reads the principle: PRINCIPLES.md:13 bars a second store ("No feature may introduce a second store (queue, socket, sidecar database)"), and a rebuildable derived index is not a store. CAPABILITIES.md:43-46 positively sanctions "a derived knowledge-graph index" as a family-1 memory-store concern, and RFC-007 sequences the persisted graph.json as Phase 6 DEFERRED, not rejected.

This is the code-side mirror of the RFC correction that shipped in #586 (8808dd7), whose commit message recorded the obligation, including that the correct Principle-1 bar on an external graph DB must be preserved.

What changed

scripts/em-graph.mjs header comment, 3 lines to 10. It now:

• keeps the Principle-1 bar, correctly scoped to an external graph DB (Neo4j, Memgraph);
• names the honest v1 rationale, the freshness-versus-rebuild-cost trade: always current with zero staleness surface, at the cost of rebuilding on every call;
• records that a persisted graph.json would be a rebuildable derived index, not a second store, admissible and deferred by RFC-007 to Phase 6 with the trigger condition named, not ruled out.

docs/rfcs/RFC-007-graph-projection.md, three passages that asserted the comment was still uncorrected. Left unfixed, the RFC would have become false the moment this landed, which is the same doc/code divergence #581 existed to close:

• :218 and :220/:251 no longer carry the em-graph.mjs:29-33 / :29-31 line anchors into a comment block, since that is the decay class this PR cleans up;
• :220 and :251 flip from "the code comment is to be corrected in a later phase" to the corrected state;
• :220's derived-index count is corrected from four to five: tokens.json was missing (scripts/em-rebuild-index.mjs:317-318, and enumerated as derivedNames alongside the other four at scripts/em-backup.mjs:1740). Issue em-graph.mjs:29-31 cites Principle 1 to reject a persisted derived index (over-read) #587's own body carried the same undercount.

docs/plans/issue-587-em-graph-principle-comment.md, the frozen build spec, committed with the slice per repo convention.

Deliberately not changed

• docs/EM_SCRIPTS_GUIDE.md:734-735 cites no principle and its "no sidecar DB" claim remains true, so it is not same-class and not drift.
• docs/plans/RFC-007-P2-rule-rfc-nodes.md:330 quotes the old comment as a record of a merged slice. Shipped plan docs are history and are not rewritten.
• No Implementation-ledger row at :443-447. A comment fix is not a phase, and a row would need a merge SHA, i.e. a placeholder plus a follow-up PR, which is exactly the Two pre-existing RFC ledger rows cite feature-branch tip SHAs instead of main-merge commits #582/Fill the RFC-007 Phase 2 ledger provenance from git, not a placeholder #605 pattern. The :220 rewrite describes the corrected state directly and needs no SHA.

Verification

All run by the orchestrator on disk, not accepted from a seat report.

Claim	Command	Observed
Behavior unchanged	node tests/test-em-graph.mjs	10 passed, 0 failed
Phase-2 suite unchanged	node tests/test-em-graph-nodes.mjs	28 passed, 0 failed
RFC set consistent	node scripts/em-rfc-validate.mjs	✓ RFC registry consistent: 14 RFCs in _index.json, 14 canonical files, 14 README table rows.
Comment does not lie about its own module	grep -cE "writeFileSync|renameSync|mkdirSync|appendFileSync|writeFile\(" scripts/em-graph.mjs	0
Old citation gone	grep -c "graph DB (Principle 1)" scripts/em-graph.mjs	0
Correct bar retained	grep -c "barred by Principle 1" scripts/em-graph.mjs	1
Decayed anchors gone	grep -c "em-graph.mjs:29-31" docs/rfcs/RFC-007-graph-projection.md	0
Live invocation works	node scripts/em-graph.mjs --hubs --top 3 --scope local	valid JSON, "status":"ok", exit 0

A repo-wide same-class sweep over 83 entries under scripts/, all 25 RFCs, 30 plans, README.md, CAPABILITIES.md, PRINCIPLES.md, CLAUDE.md, instructions/, and tests/ found this to be the only site in the repo that over-read Principle 1 against a derived index. Every other Principle-1 citation correctly targets a genuine second store, and no Principle number in scripts/ is stale relative to PRINCIPLES.md. No test pins the comment text, so the reword cannot produce a hollow green.

Review

Adversarial frozen-diff review by pi --provider kimi-coding --model k3: ACCEPT, 0 blockers, 0 minors, 2 nits. It verified each clause of the new comment against PRINCIPLES.md, CAPABILITIES.md, and RFC-007 with captured output, confirmed the U+2014 byte sequence by hexdump and line widths by awk, independently confirmed the five-index count, and ruled ACCEPT on both flagged deviations (naming documents and sections instead of line anchors; no ledger row).

Nit dispositions: the RFC-internal SQLite attribution split (:18/:335 under Principle 1 versus :364 under the zero-deps pin) is pre-existing, not introduced here, and is filed as a comment on #606. The "tracked" past tense at :220 is rejected as no-change, since the RFC is read in its merged state once #587 is closed.

Follow-up filed

#606 — RFC-007:259-262 cites the Alternatives table for Phase 4 and Phase 6. Phase 4 is at :385 and Phase 6 at :387, while :361 is the Alternatives table header and :363 is the external-graph-DB row. Issue #587's body and workplan v265 both inherited that bad anchor. Filed with all 5 DEFER fields.

[lantiscooperdev]

Tiered multi-agent pipeline — issue #587 slice, round 1 ACCEPT

Adversarial frozen-diff review of the Principle-1 comment correction, plus the RFC-side passages that would have gone stale the moment the code landed.

Verdict trail

Round	Scope	Reviewer	Verdict	Artifact
Scout ×3	code surface, governing-artifact grounding, repo-wide same-class sweep	Claude Sonnet (read-only)	maps returned, all claims orchestrator-ground-truthed	in-session
1	frozen diff vs frozen spec, 6 named attack axes	pi --provider kimi-coding --model k3	ACCEPT — 0 blockers, 0 minors, 2 nits	docs/plans/issue-587-em-graph-principle-comment.md §19

Review seat was directed by the operator in-session, replacing the standing GLM/neuralwatt reviewer.

What the review actually verified

Clause-by-clause, each with captured output rather than assertion:

Clause of the new comment	Verdict	Evidence
"an external graph DB (Neo4j, Memgraph) is barred by Principle 1, which admits no second store"	TRUE	PRINCIPLES.md:13 parenthetical; RFC-007:363
"built fresh from index rows + one body pass when cites is selected"	TRUE	em-graph.mjs cites-gated body pass; only fs call is a readFileSync
"always current with zero staleness surface"	TRUE	corroborated verbatim by RFC-007:166
"rebuildable derived index … admissible under Principle 1 and CAPABILITIES.md family 1"	TRUE	CAPABILITIES.md:42-46; identical wording at RFC-007:248
"deferred by RFC-007 to Phase 6 (trigger: rebuild cost per query becomes the binding constraint)"	TRUE	RFC-007:387 Phase 6 row
comment does not lie about its own module	PASS	grep -cE "writeFileSync|renameSync|mkdirSync|appendFileSync" scripts/em-graph.mjs → 0
house style preserved	PASS	hexdump confirms e2 80 94 (U+2014); awk confirms 56-77 col widths

Attack axes that came back clean: SQLite omission is correct (the comment scopes its Principle-1 bar to exactly the pair :363 scopes it to, and an omission cannot be false); "admissible" is RFC-007's own term, not an overreach; RFC-007:220 is internally coherent across its two-pass edit (tense arc formerly → analysis → applied); and a repo-wide sweep found no other site still asserting or quoting the comment as current.

Two orchestrator-side defects caught by independent re-verification

Recorded because the seat reported success on both:

1. Under-scoped spec step. The original Step 4 replaced only the trailing sentence of RFC-007:220, leaving the paragraph opening asserting in the present tense that the comment was still uncorrected, still carrying the decayed :29-31 anchor. Fixed by an amended Step 8; grep -c "currently justifies" → 0, grep -c "em-graph.mjs:29-31" → 0.
2. Seat stale-green. The builder reported verifies "1a/1b/1c/3/4a/4b/5a/5b/6a" and silently omitted 6b, the one that failed. Per the spec's §A.3 it owed a STOP block. The orchestrator's independent re-run is what caught it.

The seat did find one real bug in the spec: Verify 1d's regex included graph\.json, which the new comment legitimately contains, making the verify unsatisfiable. REQ-6 now greps only the fs call sites.

Test posture (final)

• node tests/test-em-graph.mjs → 10 passed, 0 failed (unchanged)
• node tests/test-em-graph-nodes.mjs → 28 passed, 0 failed (unchanged)
• node scripts/em-rfc-validate.mjs → 14 RFCs consistent
• node scripts/em-graph.mjs --hubs --top 3 --scope local → valid JSON, "status":"ok", exit 0
• CI: 8/8 checks green

No test was added. Nothing new is behaviorally testable, and a test asserting comment text would pin prose and make future honest rewording a failure. Stated so the absence is a decision, not an oversight.

Follow-ups filed (Rule 18 step 9)

• #606 — RFC-007:259-262 cites the Alternatives table for Phase 4 and Phase 6 (real rows are :385 and :387; :363 is the external-graph-DB row). Issue #587's own body and workplan v265 both inherited that bad anchor. All 5 DEFER fields.
• #606 comment — nit 1 from this round: RFC-007 attributes the SQLite rejection to Principle 1 at :18/:335 but to the zero-deps pin at :364. Pre-existing at 2df117f, not introduced here.
• Nit 2 ("Issue #587 tracked that code-side fix" past tense) dispositioned REJECT, no change: the RFC is read in its merged state, at which point #587 is closed and the tense is correct.

What user approval is checking

Per Rule 17, your approval in the UI is the review-required gate. Substantive review is complete and closed ACCEPT with zero blockers. The judgement calls worth your eye are the two accepted deviations: naming documents and sections instead of the line anchors #587 proposed, and adding no Implementation-ledger row. Please do a final read of the diff and approve, or HOLD with findings.

🤖 Posted by lantiscooperdev (bot account) — Claude Opus 5 (1M context).